clanker-hackathon / engine /pendulum.py
deucebucket's picture
clanker hackathon gradio entry (verification copy)
ecca2a3 verified
Raw
History Blame Contribute Delete
67 kB
"""Fixed Physics Layer — computes VADUGWI from structural analysis.
Pipeline: text -> classify words -> compute proximity -> detect structures -> apply physics -> VADUGWI
The physics (momentum, force application, blending) are FIXED.
The inputs come from the structural layers.
Refactored into a modular pipeline: each stage is an independent function
that reads from and writes to a shared context dict. Stages can be chained,
bypassed, or swapped.
"""
import re
from math import tanh, exp, log
from typing import List, Optional, Tuple
from .shared import VADUG, PersonalityVector
from .word_classifier import WordRole, classify_sentence, _clean
from .proximity import proximity_coefficient
from .vocabulary import VOCABULARY
from .structures import StructureDetector, StructureMatch
from .force_flow import resolve_force_flow, compute_flow_modifiers, compute_intent
from .phase import is_solvent, get_phase
# ── Physics constants (fixed, never tuned per-sentence) ─────────
CENTER = 128.0
M_BASE = 0.557 # champion v2: genetically optimized 2026-04-03
M_AROUSAL_SCALE = 0.25 # arousal-scaled momentum: high A = sticky state
M_NEGATIVITY_BIAS = 1.15 # negative states are stickier than positive
M_POSITIVITY_EASE = 0.90 # positive transitions are easier
M_MIN = 0.30 # floor: never fully unresponsive to input
M_MAX = 0.95 # ceiling: never fully locked in state
FORCE_SCALE = 1.405 # champion v2
DIRECT_PUSH_CAP = 1.0 # champion v2: max push
DIRECT_PUSH_TRIGGER = 86.2 # champion v2
SATURATION = 120.0 # tanh saturation: smooth compression replaces hard clamp
# Mundane dampening: massless context atoms absorb crisis energy
# D = (G_t + ε) / (G_t + α * |dV|)
# Gemini's improvement: ε decays exponentially with |dV| so strong crisis
# words near zero-gravity subjects get crushed almost to zero.
# ε = e^(-λ|dV|) → at |dV|=35: ε=0.03, at |dV|=60: ε=0.002
MUNDANE_ALPHA = 0.04 # sensitivity: how much dV matters
MUNDANE_EPSILON = 1.0 # floor: avoid division by zero (council suggested lower but ecosystem is tuned to this)
MUNDANE_DV_THRESHOLD = 25 # only dampen high-charge atoms (crisis words)
# ── Compound resolution tables ──────────────────────────────────
_BOOKEND_COMPOUNDS = {
"shut": "up", # shut [anything] up = silence command
"get": "out", # get [the fuck] out = expulsion command
"fuck": "off", # fuck [right] off = rejection command
"back": "off", # back [the hell] off = distance command
"piss": "off", # piss off = rejection
}
_SPICE_WORDS = {"the", "a", "an", "fuck", "fucking", "fuckin", "damn",
"god", "hell", "mother", "motherfuckin", "stupid",
"bitch", "ass", "right", "just", "already", "up",
"freakin", "freaking", "effing"}
_ONE_AS_QUANTIFIER = {"thing", "person", "place", "way", "reason", "time",
"day", "moment", "word", "chance", "step", "bit"}
_CONTINUATION_PAIRS = {
("wont", "stop"), ("wont", "quit"), ("wont", "end"),
("wont", "leave"), ("wont", "go"),
("cant", "stop"), ("cant", "quit"),
("dont", "stop"), ("doesnt", "stop"),
("never", "stop"), ("never", "stops"), ("never", "end"),
("never", "ends"), ("never", "quit"),
}
_UNIVERSAL_ADDRESS = {
"everyone", "everybody", "anyone", "anybody", "all", "people",
"folks", "y'all", "yall", "ladies", "gentlemen",
}
_SECOND_PERSON = {"you", "your", "yours", "yourself", "yourselves"}
_ABSENCE_MARKERS = {"without", "havent", "haven't", "hasnt", "hasn't"}
_ABSENCE_FOLLOWERS = {"had", "been", "felt", "seen", "gotten", "experienced"}
_PERSON_ROLES = {"SELF_REF", "OTHER_REF", "RELATION_REF"}
_CHOICE_VERBS = {"choose", "chose", "choosing", "decide", "pick", "picked"}
# ── Trigram compound bonds (Council Round 8) ───────────────────
_COMPOUND_BONDS_TRI = {
("got", "laid", "off"): "laidoff",
("get", "laid", "off"): "laidoff",
("got", "kicked", "out"): "kickedout",
("got", "locked", "out"): "lockedout",
("got", "ripped", "off"): "rippedoff",
("got", "thrown", "out"): "thrownout",
("got", "cut", "off"): "cutoff",
("got", "burned", "out"): "burnedout",
("got", "wiped", "out"): "wipedout",
}
# ── Compound bond table (Council Round 8) ──────────────────────
# Multi-word phrases that form molecular bonds with emergent charge.
# Resolved in tokenizer → single atom → charge from vocabulary.
_COMPOUND_BONDS = {
# Negative life events
("laid", "off"): "laidoff",
("food", "poisoning"): "foodpoisoning",
("broke", "down"): "brokedown",
("locked", "out"): "lockedout",
("kicked", "out"): "kickedout",
("passed", "away"): "passedaway",
("cut", "off"): "cutoff",
("thrown", "out"): "thrownout",
("ripped", "off"): "rippedoff",
("wiped", "out"): "wipedout",
("burned", "out"): "burnedout",
("shut", "down"): "shutdown",
("backed", "out"): "backedout",
("dropped", "out"): "droppedout",
("sold", "out"): "soldout",
("stressed", "out"): "stressedout",
("freaked", "out"): "freakedout",
("ruled", "out"): "ruledout",
("washed", "out"): "washedout",
("checked", "out"): "checkedout",
# Positive resolution events
("cancer", "free"): "cancerfree",
("debt", "free"): "debtfree",
("pain", "free"): "painfree",
("pulled", "off"): "pulledoff",
("pulled", "through"): "pulledthrough",
("worked", "out"): "workedout",
("paid", "off"): "paidoff",
("turned", "around"): "turnedaround",
# Neutral/procedural compounds
("log", "in"): "login",
("sign", "up"): "signup",
("check", "in"): "checkin",
("pick", "up"): "pickup",
("set", "up"): "setup",
("WiFi", "down"): "wifidown",
("wifi", "down"): "wifidown",
# Hype/intensifier compounds (slang batch 1)
("lets", "go"): "letsgo",
("let's", "go"): "letsgo",
("as", "hell"): "ashell",
("so", "back"): "soback",
}
# "w" reads as WIN only after these or sentence-finally; otherwise it's "with"
_W_WIN_PRECEDERS = frozenset({
"a", "the", "big", "massive", "huge", "major", "another",
"certified", "what",
})
_ELONG_RUN = re.compile(r"([a-zA-Z])\1{2,}")
_TRAP_GENRE_FOLLOWERS = frozenset({
"remix", "beat", "beats", "music", "song", "songs", "mix",
"playlist", "banger", "instrumental", "album", "track", "tracks",
})
_FIRE_HAZARD_PRECEDERS = frozenset({"a", "the"})
_FIRE_HAZARD_FOLLOWERS = frozenset({
"in", "at", "on", "alarm", "department", "drill", "spread", "broke",
})
_FIRE_PRAISE_PRECEDERS = frozenset({
"is", "was", "its", "it's", "thats", "that's", "so", "straight",
"absolute", "be",
})
_FILTHY_SPORTS_NOUNS = frozenset({
"play", "move", "pitch", "dunk", "goal", "crossover", "mixtape", "beat",
})
_MID_TEMPORAL_FOLLOWERS = frozenset({
"morning", "afternoon", "evening", "day", "week", "month", "year",
"semester", "summer", "winter", "spring", "fall", "season",
})
# ── Stage 1: Tokenize ──────────────────────────────────────────
def tokenize(context: dict) -> dict:
"""Split text into words, resolve compound phrases and double negations.
Reads: context["text"]
Writes: context["words"]
"""
text = context["text"]
words = text.split()
if not words:
context["words"] = []
return context
# Bookend compounds: opener ... closer with variable filling
collapsed = list(words)
for opener, closer in _BOOKEND_COMPOUNDS.items():
start_idx = None
for idx, w in enumerate(collapsed):
if w.lower() == opener:
start_idx = idx
elif w.lower() == closer and start_idx is not None and idx - start_idx <= 8:
filling = [collapsed[j].lower() for j in range(start_idx + 1, idx)]
all_spice = all(f in _SPICE_WORDS for f in filling)
if all_spice:
compound = opener + closer
collapsed = collapsed[:start_idx] + [compound] + collapsed[idx+1:]
start_idx = None
break
words = collapsed
# Double negation compounds and special pairs
sentence_cleaned = {w.lower().rstrip(".,!?;:'\"") for w in words}
resolved = []
i = 0
context.setdefault("elongation", False)
while i < len(words):
w_low = words[i].lower()
next_low = words[i + 1].lower() if i + 1 < len(words) else ""
pair = (w_low, next_low)
# "w" sense split: WIN keeps its vocab charge; "with" carries no
# force at all (a zero-force atom would still drain V toward center)
if w_low.rstrip(".,!?;:'\"") == "w":
prev_clean = words[i - 1].lower().rstrip(".,!?;:'\"") if i > 0 else ""
if prev_clean in _W_WIN_PRECEDERS or i == len(words) - 1:
resolved.append(words[i])
i += 1
continue
# "extra" sense split: prenominal quantity ("extra cheese") carries
# no charge; the slang sense ("she's so extra") keeps its vocab force
if w_low.rstrip(".,!?;:'\"") == "extra" and next_low:
nf = VOCABULARY.get(next_low.rstrip(".,!?;:'\""))
if nf is None or abs(nf[0]) < 10:
i += 1
continue
# "trap" sense split: music genre ("trap remix") is inert; the
# entrapment sense keeps its vocab force
if (w_low.rstrip(".,!?;:'\"") == "trap"
and next_low.rstrip(".,!?;:'\"") in _TRAP_GENRE_FOLLOWERS):
i += 1
continue
# "fire" sense split: literal hazard vs predicate slang praise;
# anything else keeps the (mild) stored vocab tuple
if w_low.rstrip(".,!?;:'\"") == "fire":
prev_clean = words[i - 1].lower().rstrip(".,!?;:'\"") if i > 0 else ""
nxt_clean = next_low.rstrip(".,!?;:'\"")
if prev_clean in _FIRE_HAZARD_PRECEDERS or nxt_clean in _FIRE_HAZARD_FOLLOWERS:
resolved.append("firehazard")
i += 1
continue
if prev_clean in _FIRE_PRAISE_PRECEDERS:
resolved.append("firepraise")
i += 1
continue
# "filthy" sense split: sports/performance praise when a sports
# noun appears anywhere in the sentence
if (w_low.rstrip(".,!?;:'\"") == "filthy"
and not _FILTHY_SPORTS_NOUNS.isdisjoint(sentence_cleaned)):
resolved.append("filthypraise")
i += 1
continue
# "mid" sense split: temporal compound ("mid morning") is inert;
# the dismissive slang sense keeps its vocab force
if (w_low.rstrip(".,!?;:'\"") == "mid"
and next_low.rstrip(".,!?;:'\"") in _MID_TEMPORAL_FOLLOWERS):
i += 1
continue
if pair in _CONTINUATION_PAIRS:
has_negative_before = any(
VOCABULARY.get(words[j].lower(), (0,))[0] < -25
for j in range(max(0, i - 3), i)
)
if has_negative_before:
i += 2
else:
resolved.append(words[i])
i += 1
elif w_low == "no" and next_low == "one":
next_after = words[i + 2].lower() if i + 2 < len(words) else ""
if next_after in _ONE_AS_QUANTIFIER:
resolved.append(words[i])
i += 1
else:
resolved.append("nobody")
i += 2
elif w_low == "come" and next_low == "on":
resolved.append("comeon")
i += 2
elif w_low == "killed" and next_low == "it":
resolved.append("killedit")
i += 2
elif w_low == "goes" and next_low == "hard":
resolved.append("goeshard")
i += 2
elif w_low == "hit" and next_low == "different":
resolved.append("hitdifferent")
i += 2
elif w_low == "running" and next_low == "late":
resolved.append("runninglate")
i += 2
elif w_low == "made" and next_low == "it":
# "made it" → achievement compound
next_after = words[i + 2].lower() if i + 2 < len(words) else ""
if next_after == "through":
resolved.append("madeitthrough")
i += 3
else:
resolved.append("madeit")
i += 2
elif w_low == "came" and next_low == "back":
# Look ahead for "negative" → medical idiom (good news)
next_after = words[i + 2].lower() if i + 2 < len(words) else ""
if next_after == "negative":
resolved.append("camebacknegative")
i += 3
else:
resolved.append(words[i])
i += 1
elif w_low == "closed" and next_low == "on":
resolved.append("closedon")
i += 2
elif w_low == "no" and next_low == "cap":
resolved.append("nocap")
i += 2
elif w_low == "no" and next_low == "way":
resolved.append("noway")
i += 2
else:
# Council Round 8: compound bond resolution (bigram + trigram)
w_clean = w_low.rstrip(".,!?;:'\"")
n_clean = next_low.rstrip(".,!?;:'\"")
# Elongation collapse: "GOOOOOO" -> "go" for vocab-missing tokens
# so compound bonds and vocab lookups still land. Doubled form
# is tried first because runs hide geminates: "GOOOOOD" must
# collapse to "good", not "god"
def _alt(tok):
if tok and tok not in VOCABULARY and _ELONG_RUN.search(tok):
two = _ELONG_RUN.sub(r"\1\1", tok)
if two in VOCABULARY:
return two
return _ELONG_RUN.sub(r"\1", tok)
return tok
w_alt = _alt(w_clean)
n_alt = _alt(n_clean)
# Trigram check first
if i + 2 < len(words):
n2_clean = words[i + 2].lower().rstrip(".,!?;:'\"")
tri = (w_clean, n_clean, n2_clean)
if tri in _COMPOUND_BONDS_TRI:
resolved.append(_COMPOUND_BONDS_TRI[tri])
i += 3
continue
n2_alt = _alt(n2_clean)
tri_alt = (w_alt, n_alt, n2_alt)
if tri_alt != tri and tri_alt in _COMPOUND_BONDS_TRI:
resolved.append(_COMPOUND_BONDS_TRI[tri_alt])
context["elongation"] = True
i += 3
continue
# Bigram check
bi = (w_clean, n_clean)
bi_alt = (w_alt, n_alt)
if bi in _COMPOUND_BONDS:
resolved.append(_COMPOUND_BONDS[bi])
i += 2
elif bi_alt != bi and bi_alt in _COMPOUND_BONDS:
resolved.append(_COMPOUND_BONDS[bi_alt])
context["elongation"] = True
i += 2
elif w_alt != w_clean and w_alt in VOCABULARY:
resolved.append(w_alt)
context["elongation"] = True
i += 1
else:
resolved.append(words[i])
i += 1
words = resolved
context["words"] = words
return context
# ── Stage 2: Classify ──────────────────────────────────────────
def classify(context: dict) -> dict:
"""Classify words into structural roles, apply perspective remapping.
Reads: context["words"], context["perspective"], context.get("personality")
Writes: context["roles"], context["perspective"] (may be resolved from "auto"),
context["has_universal"]
"""
words = context["words"]
perspective = context.get("perspective", "speaker")
personality = context.get("personality")
roles = classify_sentence(words)
# Universal address detection
has_universal = any(w.lower() in _UNIVERSAL_ADDRESS for w in words)
context["has_universal"] = has_universal
# Auto-detect perspective
if perspective == "auto":
has_self = any(wr.role == "SELF_REF" for wr in roles)
has_second = any(wr.word.lower() in _SECOND_PERSON for wr in roles)
if has_self:
perspective = "speaker"
elif has_second:
perspective = "listener"
elif has_universal:
perspective = "listener"
else:
perspective = "bystander"
# Perspective remapping
if perspective == "listener":
for wr in roles:
if wr.role == "SELF_REF":
wr.role = "OTHER_REF"
elif wr.role == "OTHER_REF":
wr.role = "SELF_REF"
elif perspective == "bystander":
if has_universal:
for wr in roles:
if wr.role == "SELF_REF":
wr.role = "OTHER_REF"
else:
for wr in roles:
if wr.role in ("SELF_REF", "OTHER_REF"):
wr.role = "NEUTRAL"
# Bystander self-projection
if perspective == "bystander" and personality is not None:
bystander_w = personality.assertiveness
if bystander_w < 80:
_THIRD_PERSON = {"she", "he", "her", "him", "they", "them"}
for idx, wr in enumerate(roles):
if wr.word.lower() in _THIRD_PERSON and wr.role == "NEUTRAL":
for j in range(max(0, idx - 3), min(len(roles), idx + 4)):
if j != idx and roles[j].role == "EMOTIONAL":
w_clean = _clean(roles[j].word)
forces = VOCABULARY.get(w_clean)
if forces and forces[0] < -10:
wr.role = "SELF_REF"
break
# Absence scope
absence_scope = set()
for i, wr in enumerate(roles):
if wr.word in _ABSENCE_MARKERS:
if wr.word == "without":
for j in range(i + 1, min(i + 4, len(roles))):
if roles[j].role not in _PERSON_ROLES:
absence_scope.add(j)
elif i + 1 < len(roles) and roles[i + 1].word in _ABSENCE_FOLLOWERS:
for j in range(i + 2, min(i + 7, len(roles))):
absence_scope.add(j)
# Forced choice cancellation
forced_choice_scope = set()
for i, wr in enumerate(roles):
if wr.word == "between":
has_choice = any(roles[j].word in _CHOICE_VERBS
for j in range(max(0, i - 3), i))
if has_choice:
for j in range(i + 1, min(i + 6, len(roles))):
forced_choice_scope.add(j)
context["roles"] = roles
context["perspective"] = perspective
context["absence_scope"] = absence_scope
context["forced_choice_scope"] = forced_choice_scope
return context
# ── Stage 2.5: Interpret Context ──────────────────────────────
# GPT's insight: "words → role/context interpretation → forces → result"
# This layer reclassifies roles based on PRAGMATIC FUNCTION before
# forces are computed. The physics is the same — the interpretation
# of what role each atom plays changes based on molecular context.
_DISCOURSE_AFFIRMS = frozenset({"no", "nah", "nope"})
_DISCOURSE_LOOKAHEAD_POS = frozenset({
"good", "fine", "right", "great", "cool", "ok", "okay",
"way", "cap", "doubt", "kidding", "worries",
})
_DISCOURSE_SELF_TOKENS = frozenset({"i", "im", "i'm", "ive", "i've"})
_LIQUID_LITERAL_FOLLOWERS = {
"cooked": frozenset({"dinner", "breakfast", "lunch", "food", "meal",
"meals", "rice", "chicken", "steak", "pasta",
"eggs"}),
}
_EXPLETIVE_WORDS = frozenset({"shit", "fuck", "damn", "hell", "goddamn"})
_COUNTERFACTUAL_MARKERS = frozenset({
"supposed", "would", "should", "could", "wished", "hoped",
})
_PAST_TRUST_VERBS = frozenset({
"trusted", "believed", "thought", "assumed", "expected",
})
_INSTRUCTIONAL_CUES = frozenset({
"please", "proceed", "enter", "select", "follow",
"click", "press", "navigate", "submit", "confirm",
"objective", "quest", "mission", "instructions",
})
def interpret_context(context: dict) -> dict:
"""Reinterpret word roles based on pragmatic context.
Runs AFTER classify, BEFORE force computation. Modifies roles
and sets context flags that downstream stages use.
Fixes:
1. Discourse markers: "no we good" → "no" becomes DISCOURSE_AFFIRM
2. Expletive-as-intensifier: "shit you are right" → "shit" becomes AMPLIFIER
3. Register detection: instructional text → force dampening flag
4. Counterfactual marking: "supposed to" → positive forces inverted
5. Hard negator inversion: NEGATOR + strong positive → full sign flip
"""
roles = context["roles"]
if not roles:
return context
words = [r.word for r in roles]
n = len(roles)
# ── 1. Discourse markers ──────────────────────────────────
# "no we good" — "no" followed by positive content = affirm, not negate
for i, wr in enumerate(roles):
if wr.word in _DISCOURSE_AFFIRMS and wr.role == "NEGATOR":
# Check 3-token lookahead for positive signal
lookahead = words[i+1:i+4]
has_positive_ahead = any(w in _DISCOURSE_LOOKAHEAD_POS for w in lookahead)
# Also check: is there a positive emotional word ahead?
has_emo_pos = any(
roles[j].force and roles[j].force[0] > 10
for j in range(i+1, min(i+4, n))
)
if has_positive_ahead or has_emo_pos:
# Reclassify: this "no" is discourse, not negation
wr.role = "FILLER"
wr.base_role = "FILLER"
elif i == 0 and n > 1 and roles[1].word in _DISCOURSE_SELF_TOKENS:
# "no, i'm sick" — sentence-initial answer-marker before a
# self-statement negates the (implied) question, not the
# clause that follows
wr.role = "FILLER"
wr.base_role = "FILLER"
# ── 2. Expletive-as-intensifier ───────────────────────────
# "shit you are right" — sentence-initial expletive + positive/affirm content
_AFFIRM_WORDS = frozenset({
"right", "true", "yes", "exactly", "correct", "agreed",
"good", "great", "nice", "thanks", "thank",
})
# Rescue sites: sentence-initial, or right after a demonstrative
# ("This shit slaps" — the demonstrative marks the expletive as subject
# stand-in, not invective)
_DEMONSTRATIVES = frozenset({"this", "that", "the"})
for idx in range(n):
if roles[idx].word not in _EXPLETIVE_WORDS:
continue
if idx != 0 and roles[idx - 1].word not in _DEMONSTRATIVES:
continue
# Check if next clause has positive content OR affirmative words
pos_ahead = sum(
1 for j in range(idx + 1, min(idx + 6, n))
if (roles[j].force and roles[j].force[0] > 10)
or roles[j].word in _AFFIRM_WORDS
)
if pos_ahead > 0:
# Convert expletive to amplifier. The explicit zero tuple (not
# None) blocks every downstream `force or VOCABULARY.get(word)`
# fallback from re-resolving the expletive's negative charge
roles[idx].role = "AMPLIFIER"
roles[idx].force = (0, 0, 0, 0, 0)
# ── 2b. Degree-adverb slang ───────────────────────────────
# "stupid good" / "insanely talented" — the negative adjective is a
# degree marker on the positive word, not its own charge
_DEGREE_SLANG = frozenset({"stupid", "dumb", "insanely", "ridiculously",
"stupidly"})
for idx in range(n - 1):
if roles[idx].word not in _DEGREE_SLANG:
continue
nf = roles[idx + 1].force or VOCABULARY.get(roles[idx + 1].word)
if nf and nf[0] >= 20:
roles[idx].role = "AMPLIFIER"
roles[idx].force = (0, 0, 0, 0, 0)
# ── 3. Register detection (Council Round 7) ────────────────
# CONVERSATIONAL = full force (speaker emitting charge)
# LITERARY = dampened force (charge is described/reported, not felt)
# EXPOSITORY = heavily dampened (procedural/instructional text)
#
# Detection via Dielectric Index (Gemini) + agency structure (GPT):
# High article/3rd-person density + no 1st/2nd person + no casual markers = narration
# The medium is dense — weak atoms scatter, strong atoms punch through dampened.
instructional_count = sum(1 for w in words if w in _INSTRUCTIONAL_CUES)
_CASUAL_SIGNALS = {"im", "ive", "youre", "dont", "cant", "wont",
"gonna", "wanna", "gotta", "lol", "lmao", "bruh", "bro",
"dude", "omg", "tbh", "ngl", "fr", "nah", "yeah", "yep",
"hey", "yo", "haha", "ok", "okay"}
_ARTICLES = {"the", "a", "an", "this", "these", "those"}
_THIRD_PERSON = {"he", "she", "they", "it", "his", "her", "its", "their",
"him", "them"}
_FIRST_PERSON = {"i", "im", "ive", "my", "me", "we", "us", "our"}
_SECOND_PERSON = {"you", "your", "youre"}
_LITERARY_VERBS = {"seized", "seizing", "walked", "strode", "gazed",
"muttered", "exclaimed", "remarked", "observed",
"whispered", "cried", "replied", "declared"}
casual_count = sum(1 for w in words if w in _CASUAL_SIGNALS)
has_exclamation = any(w.endswith('!') for w in words)
has_quotes = context["text"].count('"') >= 2
c_art = sum(1 for w in words if w in _ARTICLES)
c_3rd = sum(1 for w in words if w in _THIRD_PERSON)
c_1st = sum(1 for w in words if w in _FIRST_PERSON)
c_2nd = sum(1 for w in words if w in _SECOND_PERSON)
c_lit_verb = sum(1 for w in words if w in _LITERARY_VERBS)
past_count = sum(1 for w in words if w.endswith("ed") or w in {"was", "were", "had", "been"})
# Dielectric Index: (articles + 3rd person) / total tokens
dx = (c_art + c_3rd) / max(1, n)
past_ratio = past_count / max(1, n)
# Observation score (GPT): agency-based
obs_score = dx + 0.3 * past_ratio + (0.1 if casual_count == 0 else 0)
# Literary verb bonus (Claude)
if c_lit_verb >= 1:
obs_score += 0.2
# Decision hierarchy:
# 1. Instructional cues → EXPOSITORY (strongest override)
# 2. Casual markers or exclamation or quotes → CONVERSATIONAL
# 3. High observation score + no 1st/2nd person + length >= 5 → LITERARY
# 4. Default → CONVERSATIONAL
if instructional_count >= 1:
register = "EXPOSITORY"
context["register_dampener"] = 0.35
elif casual_count >= 1 or has_exclamation or has_quotes:
register = "CONVERSATIONAL"
context["register_dampener"] = 1.0
elif obs_score >= 0.50 and c_1st == 0 and c_2nd == 0 and casual_count == 0 and n >= 10:
register = "LITERARY"
# Gemini's mass-dependent scattering: applied per-word in accumulate_forces
context["register_dampener"] = 0.55
else:
register = "CONVERSATIONAL"
context["register_dampener"] = 1.0
context["register"] = register
# ── 4. Counterfactual marking ─────────────────────────────
# "supposed to", "would have" → flag for force inversion
has_counterfactual = any(w in _COUNTERFACTUAL_MARKERS for w in words)
has_past_trust = any(w in _PAST_TRUST_VERBS for w in words)
context["counterfactual"] = has_counterfactual
context["past_trust"] = has_past_trust
# ── 5. Hard negator inversion ─────────────────────────────
# NEGATOR within 3 tokens of positive EMOTIONAL (dV > 25) → flip sign
# Also reclassify the NEGATOR to FILLER so proximity doesn't double-negate
negators_consumed = set()
for i, wr in enumerate(roles):
if wr.role == "NEGATOR":
for j in range(i+1, min(n, i+4)):
jr = roles[j]
f = jr.force or VOCABULARY.get(jr.word)
if f and f[0] > 25:
# Full inversion: flip the positive word's dV
old_f = f
jr.force = (-old_f[0], old_f[1], -old_f[2], old_f[3], old_f[4])
negators_consumed.add(i)
break
# Also consume negators that precede a communication/action verb
# before distant emotional content.
# "nobody tells you grief" → "nobody" negates "tells", not "grief"
# "nobody loves me" → "nobody" SHOULD negate "loves" → don't consume
# Key: consume only when intervening word is a low-charge comm verb
_COMM_VERBS = frozenset({
"tells", "told", "says", "said", "asks", "asked",
"knows", "knew", "thinks", "thought", "warns", "warned",
"prepares", "prepared", "expects", "expected",
"mentions", "mentioned", "explains", "explained",
})
if i not in negators_consumed:
for j in range(i+1, min(n, i+3)):
if roles[j].word in _COMM_VERBS:
negators_consumed.add(i)
break
# Consumed negators become FILLER so proximity doesn't double-apply
for i in negators_consumed:
roles[i].role = "FILLER"
# ── 6. SOLVENT dissolution ────────────────────────────────
# SOLVENT words (bruh, lol, lmao, dude, etc.) dissolve LIQUID atoms
# "bruh im shook" → SOLVENT(bruh) flips LIQUID(shook) from negative to positive
# "bruh he got murdered" → SOLVENT can't dissolve SOLID(murdered)
# Literal-verb guard: a LIQUID slang verb followed by its concrete
# object ("cooked dinner") or preceded by "home" is the literal sense
# — neutral, not the stored slang-doom charge, and not dissolvable
for idx, wr in enumerate(roles):
followers = _LIQUID_LITERAL_FOLLOWERS.get(wr.word)
if not followers:
continue
nxt = roles[idx + 1].word if idx + 1 < n else ""
prev = roles[idx - 1].word if idx > 0 else ""
if nxt in followers or prev == "home":
wr.force = (0, 0, 0, 0, 0)
has_solvent = any(is_solvent(wr.word) for wr in roles)
if has_solvent:
for wr in roles:
if get_phase(wr.word) == "LIQUID":
# Get the force (may be on wr.force or in VOCABULARY)
f = wr.force or VOCABULARY.get(wr.word)
if f and f[0] < -5:
# Flip dV sign, keep arousal (the energy stays, charge flips)
flipped = (-f[0], f[1], abs(f[2]), f[3], abs(f[4]))
wr.force = flipped
context["solvent_active"] = True
else:
context["solvent_active"] = False
# ── 7. Sarcasm inversion field (Council Round 6) ────────────
# Genuine enthusiasm radiates amplification energy. Sarcasm is a
# cold molecule — positive surface with zero kinetic energy.
# Ironic onset + tepid positive + zero amplifiers → invert.
_IRONIC_ONSETS = frozenset({
"clearly", "oh", "wow", "sure", "right", "great",
"nice", "yeah", "gee", "wonderful", "brilliant", "lovely",
})
_AMPLIFIERS = frozenset({
"so", "really", "very", "super", "extremely", "absolutely",
"honestly", "seriously", "genuinely", "totally",
})
first_word = words[0] if words else ""
two_word = " ".join(words[:2]) if len(words) >= 2 else ""
has_ironic_onset = first_word in _IRONIC_ONSETS or two_word in ("what a", "oh great", "oh cool", "oh nice")
has_amplifier = any(w in _AMPLIFIERS for w in words) or any(w.endswith("!") for w in words)
if has_ironic_onset and not has_amplifier and not has_solvent:
# Count all charged atoms
total_charge = 0
pos_count = 0
neg_count = 0
for wr in roles:
f = wr.force or VOCABULARY.get(wr.word)
if f:
total_charge += f[0]
if f[0] > 0:
pos_count += 1
elif f[0] < 0:
neg_count += 1
# Flat affect: ironic onset + mostly near-zero/mildly negative atoms
# + no strong positive to anchor genuine enthusiasm
# This catches: "clearly this was well thought out" (all atoms -5 or 0)
# "oh cool cant wait for that" (all atoms -5 to -10)
if neg_count >= pos_count and total_charge < 0 and neg_count >= 1:
# All atoms weakly negative + ironic onset = sarcasm
# Apply a structural V penalty
context["sarcasm_inversion"] = True
context["sarcasm_penalty"] = -15.0 # applied in apply_structures
# ── 7b. Contrast sarcasm: strong positive + negative in same sentence ──
# "I am just overjoyed to clean up your mess" — "overjoyed" near "mess"
# "Thanks for that incredibly useless advice" — "thanks" near "useless"
# The CONTRAST between positive and negative in a short span = sarcasm
if not context.get("sarcasm_inversion") and not has_solvent:
strong_pos = []
strong_neg = []
for wr in roles:
f = wr.force or VOCABULARY.get(wr.word)
if f:
if f[0] >= 15:
strong_pos.append(wr.word)
elif f[0] <= -15:
strong_neg.append(wr.word)
# Both strong positive AND strong negative = contrast sarcasm
if strong_pos and strong_neg and len(words) <= 15:
context["sarcasm_inversion"] = True
context["sarcasm_penalty"] = -12.0
context["roles"] = roles
return context
# ── Stage 3: Compute Coefficients (structure detection + force flow) ─
def compute_coefficients(context: dict) -> dict:
"""Detect structures and resolve force flow.
Reads: context["roles"]
Writes: context["structures"], context["force_flow"], context["flow_mods"]
"""
roles = context["roles"]
detector = StructureDetector()
structures = detector.detect_all(roles)
force_flow = resolve_force_flow(roles)
flow_mods = compute_flow_modifiers(force_flow)
context["structures"] = structures
context["force_flow"] = force_flow
context["flow_mods"] = flow_mods
return context
# ── Stage 4: Accumulate Forces ──────────────────────────────────
def accumulate_forces(context: dict) -> dict:
"""Per-word force application loop with adaptive momentum.
Reads: context["roles"], context["absence_scope"], context["forced_choice_scope"],
context["force_flow"], context["flow_mods"],
context["register_dampener"], context["counterfactual"], context["past_trust"]
Writes: context["state_v"], context["state_a"], context["state_d"],
context["state_u"], context["state_g"], context["state_w"],
context["trace_entries"]
"""
roles = context["roles"]
absence_scope = context.get("absence_scope", set())
forced_choice_scope = context.get("forced_choice_scope", set())
force_flow = context.get("force_flow")
flow_mods = context.get("flow_mods", {})
state_v = CENTER
state_a = CENTER
state_d = CENTER
state_u = 0.0
state_g = CENTER
state_w = CENTER
trace_entries: List[dict] = []
for i, wr in enumerate(roles):
if wr.role == "POSSESSION":
vf = VOCABULARY.get(wr.word)
if vf:
word_force = (0, 0, 0, 0, max(5, vf[4]))
else:
word_force = (0, 0, 0, 0, 5)
else:
word_force = wr.force
if word_force is None:
word_force = VOCABULARY.get(wr.word)
if word_force is None:
from .fuzzy import fuzzy_match
matched = fuzzy_match(wr.word)
if matched:
word_force = VOCABULARY.get(matched)
# A zero-charge tuple on a connector/temporal/amplifier word
# carries no information — treating it as a hit would drain
# accumulated deviation back toward center, punishing sentences
# for containing function words. Only above center: positive
# deviation is protected from erosion, while below center the
# drain keeps safe text neutral (mirrors M_NEGATIVITY_BIAS /
# M_POSITIVITY_EASE asymmetry).
_INERT_ZERO_ROLES = {"CONNECTOR", "TEMPORAL", "FILLER",
"AMPLIFIER"}
if (word_force is not None
and state_v >= CENTER
and wr.role in _INERT_ZERO_ROLES
and word_force[0] == 0 and word_force[1] == 0
and word_force[2] == 0 and word_force[3] == 0
and abs(word_force[4]) <= 5):
word_force = None
# Same protection for near-zero register markers: a trailing
# "ngl"/"fr"/"tbh" is punctuation-grade, not a counter-force
if (word_force is not None
and state_v >= CENTER
and (wr.role == "REGISTER_CASUAL" or is_solvent(wr.word))
and abs(word_force[0]) <= 5 and abs(word_force[1]) <= 5
and abs(word_force[2]) <= 5 and word_force[3] == 0
and abs(word_force[4]) <= 5):
word_force = None
if word_force is None:
trace_entries.append({
"word": wr.word,
"role": wr.role,
"coeff": 0.0,
"v": round(state_v),
"a": round(state_a),
"d": round(state_d),
"u": round(state_u),
"g": round(state_g),
"w": round(state_w),
})
continue
dv, da, dd, du, dg = word_force
# ── REGISTER DAMPENING ──
# LITERARY: Gemini's mass-dependent scattering — weak atoms scatter,
# strong atoms punch through dampened. dV_eff = dV * (1 - e^(-k|dV|)) * ε
# EXPOSITORY: flat dampening (instructional text)
reg_damp = context.get("register_dampener", 1.0)
if reg_damp < 1.0:
if context.get("register") == "LITERARY":
# Mass-dependent scattering: weak atoms crushed, strong survive
scatter = 1.0 - exp(-0.05 * abs(dv)) # approaches 1.0 for strong atoms
dv = int(dv * scatter * reg_damp)
da = int(da * scatter * reg_damp)
dd = int(dd * scatter * reg_damp)
else:
# Flat dampening for EXPOSITORY
dv = int(dv * reg_damp)
da = int(da * reg_damp)
dd = int(dd * reg_damp)
# ── COUNTERFACTUAL INVERSION: positive in past/counterfactual → grief ──
if context.get("counterfactual") and dv > 10:
dv = int(dv * -0.75) # 75% inversion
elif context.get("past_trust") and dv > 5:
dv = int(dv * 0.5) # dampen positive (broken trust context)
# ── MUNDANE DAMPENING: massless context absorbs crisis energy ──
# High-charge emotional atoms (|dV| > threshold) get dampened when
# the sentence's subject/agent is a mundane noun (low gravity).
# "homework is killing me" — subject=homework (G=0) → DAMPEN
# "i want to kill myself" — subject=I (SELF_REF) → PRESERVE
# The subject is the first substantive noun before the crisis verb,
# skipping connectors, determiners, and filler.
if abs(dv) >= MUNDANE_DV_THRESHOLD and wr.role == "EMOTIONAL":
_AGENTIC_ROLES = {"SELF_REF", "OTHER_REF", "RELATION_REF"}
_SKIP_ROLES = {"CONNECTOR", "NEUTRAL", "FILLER", "TEMPORAL",
"AMPLIFIER", "NEGATOR", "COMPRESSOR", "HEDGE"}
# Scan backward for the first substantive word (subject)
agent_g = None
agent_is_person = False
for j in range(i - 1, max(-1, i - 8), -1):
jr = roles[j]
if jr.role in _AGENTIC_ROLES:
agent_is_person = True
break
# Skip function words — look for the actual subject noun
if jr.role in _SKIP_ROLES:
jf = jr.force or VOCABULARY.get(jr.word)
# Word with no force at all = maximally mundane (G=0)
if jf is None:
agent_g = 0
break
# Word with low charge + low gravity = mundane subject
if abs(jf[0]) < 10 and abs(jf[4]) < 15:
agent_g = abs(jf[4])
break
continue
if jr.role in ("EMOTIONAL", "POSSESSION"):
jf = jr.force or VOCABULARY.get(jr.word)
jg = abs(jf[4]) if jf else 0
agent_g = jg
break
# Only dampen if subject is mundane (not a person, low G)
if not agent_is_person and agent_g is not None and agent_g < 15:
D = (agent_g + MUNDANE_EPSILON) / (agent_g + MUNDANE_ALPHA * abs(dv))
dv = int(dv * D)
da = int(da * D)
# Forced choice cancellation
if i in forced_choice_scope and dv > 0:
dv = -dv
# Absence scope dampening
if i in absence_scope and dv < -10:
dv = int(dv * 0.2)
da = int(da * 0.3)
dd = int(dd * 0.3)
# "Without" as pure operator
if wr.word in ("without",) and any(j in absence_scope for j in range(i+1, min(i+4, len(roles)))):
dv = 0
da = 0
dd = 0
# Force flow direction modifiers
if i == (force_flow.force_idx if force_flow else -1):
dv = int(dv * flow_mods["v_mod"])
dd = int(dd * flow_mods["d_mod"])
coeff = proximity_coefficient(roles, i)
# Target = center + force * coefficient * scale
target_v = CENTER + dv * coeff * FORCE_SCALE
target_a = CENTER + da * coeff * FORCE_SCALE
target_d = CENTER + dd * coeff * FORCE_SCALE
target_u = du * abs(coeff) * FORCE_SCALE
target_g = CENTER + dg * coeff * FORCE_SCALE
# Direct push for strong forces
total_force = abs(dv) + abs(da) + abs(dd) + abs(du) + abs(dg)
push_strength = min(1.0, total_force / DIRECT_PUSH_TRIGGER) * DIRECT_PUSH_CAP
push_v = push_strength * (1.0 if dv * coeff >= 0 else -1.0) * abs(dv) * FORCE_SCALE
push_a = push_strength * (1.0 if da * coeff >= 0 else -1.0) * abs(da) * FORCE_SCALE
push_d = push_strength * (1.0 if dd * coeff >= 0 else -1.0) * abs(dd) * FORCE_SCALE
push_u = push_strength * abs(du) * FORCE_SCALE
push_g = push_strength * (1.0 if dg * coeff >= 0 else -1.0) * abs(dg) * FORCE_SCALE
# Adaptive momentum
m_eff = M_BASE + (state_a - CENTER) / 255.0 * M_AROUSAL_SCALE
if state_v < CENTER and target_v > state_v:
m_v = max(M_MIN, min(M_MAX, m_eff * M_NEGATIVITY_BIAS))
elif state_v > CENTER and target_v < state_v:
m_v = max(M_MIN, min(M_MAX, m_eff * M_POSITIVITY_EASE))
else:
m_v = max(M_MIN, min(M_MAX, m_eff))
m_eff = max(M_MIN, min(M_MAX, m_eff))
inv_m_v = 1.0 - m_v
inv_m = 1.0 - m_eff
inv_m_base = 1.0 - M_BASE
state_v = state_v * m_v + target_v * inv_m_v + push_v
state_a = state_a * m_eff + target_a * inv_m + push_a
state_d = state_d * m_eff + target_d * inv_m + push_d
state_u = state_u * M_BASE + target_u * inv_m_base + push_u
state_g = state_g * M_BASE + target_g * inv_m_base + push_g
# W (self-worth): valence routed through attribution (Board 2).
# R answers "how much of this force's valence is about MY worth?"
# At the force word, R comes from the force-flow arc when it resolved
# real entity tokens:
# self-declarative ("i am worthless") R = w_mod (1.5 neg / 0.7 pos)
# force aimed at self ("he told me im nothing") R = w_mod (1.2 neg / 0.9 pos)
# self harms other ("i hurt her") — guilt R = 0.43 (x0.7 damp ≈ 0.30)
# other-directed / atmospheric, no self token R = 0 (W untouched —
# "the weather is awful" says nothing about MY worth)
# Elsewhere (and when the flow resolved nothing or only implied
# entities) W falls back to the legacy SELF_REF-proximity gate.
self_ref_nearby = any(
roles[j].role == "SELF_REF" and abs(j - i) <= 4
for j in range(max(0, i - 4), min(len(roles), i + 5))
if j != i
)
w_damp = 0.7
w_flow = flow_mods["w_mod"] if force_flow and i == force_flow.force_idx else 1.0
w_apply = self_ref_nearby
if force_flow is not None and i == force_flow.force_idx and dv != 0:
_a_role = force_flow.actor_role
_t_role = force_flow.target_role
_actor_token = force_flow.actor_idx >= 0
_target_token = force_flow.target_idx >= 0
_self_token_involved = ((_a_role == "SELF_REF" and _actor_token)
or (_t_role == "SELF_REF" and _target_token))
if (_a_role == "SELF_REF" and _actor_token
and _t_role in ("OTHER_REF", "RELATION_REF") and _target_token
and dv < 0):
# Guilt: self as actor of a negative act toward another.
# Partial self-attribution — your worth dips, but the
# valence belongs mostly to what you did, not what you are.
w_flow = 0.43
w_apply = True
elif _self_token_involved:
# Self-declarative or self-targeted force: route the
# valence into W even when the SELF token sits outside
# the 4-word proximity window.
w_apply = True
if w_apply and dv != 0:
w_effective = dv * coeff * FORCE_SCALE * w_damp * w_flow
target_w = CENTER + w_effective
push_w = push_strength * (1.0 if dv * coeff >= 0 else -1.0) * abs(dv) * FORCE_SCALE * w_damp * w_flow
state_w = state_w * m_eff + target_w * inv_m + push_w
trace_entries.append({
"word": wr.word,
"role": wr.role,
"coeff": round(coeff, 3),
"v": round(state_v),
"a": round(state_a),
"d": round(state_d),
"u": round(state_u),
"g": round(state_g),
"w": round(state_w),
})
# ── BIDIRECTIONAL CORRECTION (sentence-level A+B=C) ─────────────
# Instead of a full backward pass, pre-scan for the strongest emotional
# atom. If it's in the second half of the sentence AND the forward pass
# didn't reach its polarity, apply a correction push.
# This fixes "I just got laid off from work" where the event "laidoff"
# is in the middle but momentum recovery from trailing neutral words
# erases its charge.
strongest_dv = 0
strongest_pos = 0
mid = len(roles) // 2
for i, wr in enumerate(roles):
f = wr.force or VOCABULARY.get(wr.word)
if f and abs(f[0]) > abs(strongest_dv):
strongest_dv = f[0]
strongest_pos = i
# If strongest atom is past midpoint and forward V disagrees with its polarity
if strongest_pos >= mid and abs(strongest_dv) >= 20:
fwd_dev = state_v - CENTER
atom_direction = 1 if strongest_dv > 0 else -1
fwd_direction = 1 if fwd_dev > 0 else -1 if fwd_dev < 0 else 0
if atom_direction != fwd_direction or abs(fwd_dev) < abs(strongest_dv) * 0.3:
# Forward pass didn't capture the event's polarity — apply correction
correction = strongest_dv * 0.3 * FORCE_SCALE
state_v += correction
context["state_v"] = state_v
context["state_a"] = state_a
context["state_d"] = state_d
context["state_u"] = state_u
context["state_g"] = state_g
context["state_w"] = state_w
context["trace_entries"] = trace_entries
return context
# ── Stage 5: Apply Structures ──────────────────────────────────
def apply_structures(context: dict) -> dict:
"""Apply structure detection adjustments to state.
Reads: context["state_*"], context["structures"], context["roles"]
Writes: context["state_v"], context["state_d"], context["state_u"],
context["state_g"], context["state_w"]
"""
structures = context.get("structures", [])
roles = context.get("roles", [])
state_v = context["state_v"]
state_a = context["state_a"]
state_d = context["state_d"]
state_u = context["state_u"]
state_g = context["state_g"]
state_w = context["state_w"]
for sm in structures:
if sm.pattern == "SLANG_DEATH_HUMOR":
# Nullify death word's negative dV and add back a positive-scaled version.
# The death word is NOT functioning as death -- it's an intensifier.
# V_corrected = V_raw - dV("dead") + 0.7 * abs(dV("dead"))
death_dv_total = 0
for idx in sm.matched_indices:
if idx < len(roles):
w = roles[idx].word
vf = VOCABULARY.get(w)
if vf and vf[0] < -10: # negative death word
death_dv_total += vf[0]
if death_dv_total < 0:
# Subtract the death word's accumulated negative push and add positive version.
# Use 1.2x on the nullification to account for momentum/push amplification
# during accumulation, and 0.7x for the positive reinterpretation.
correction = 1.2 * abs(death_dv_total) * FORCE_SCALE + 0.7 * abs(death_dv_total) * FORCE_SCALE
state_v += correction * sm.confidence
else:
# Fallback: pull toward center
distance = CENTER - state_v
state_v += distance * 1.2 * sm.confidence
state_w = max(state_w, CENTER)
elif sm.pattern == "AMBIGUITY_HOLD":
# Extreme V contradiction with no disambiguator: pull V toward W (neutral baseline).
# V_final = V + (W - V) * 0.85
state_v = state_v + (state_w - state_v) * 0.85 * sm.confidence
elif sm.pattern == "RECOVERY_MILESTONE":
# Recovery milestone: apply v_weight as direct positive boost
state_v += sm.v_weight * sm.confidence * FORCE_SCALE
elif sm.pattern == "HEDGED_ASSESSMENT":
# A trailing hedge ("i guess", "i suppose") DAMPS conviction;
# it does not add negative content. The hedge word's raw vocab
# force (e.g. "guess" V-21) already accumulated as if it were
# content -- refund it (1.3x for momentum amplification), then
# apply the structure's mild damp.
hedge_dv = 0.0
hedge_dw = 0.0
for idx in sm.matched_indices:
if idx < len(roles):
hf = roles[idx].force or VOCABULARY.get(roles[idx].word)
if hf:
if hf[0] < 0:
hedge_dv += hf[0]
if len(hf) > 4 and hf[4] < 0:
hedge_dw += hf[4]
if hedge_dv < 0:
state_v += 1.3 * abs(hedge_dv) * FORCE_SCALE
if hedge_dw < 0:
state_w += 1.3 * abs(hedge_dw) * FORCE_SCALE
state_v += sm.v_weight * sm.confidence * FORCE_SCALE
elif sm.pattern in ("SARCASM_INVERSION", "BRAVADO", "DIRECTED_POSITIVE", "EXCLUDED_POSITIVE", "GRIEF_LOSS", "ATMOSPHERIC_GRIEF", "RHETORICAL_SELF_NEGATION", "REPORTED_COMFORT", "PASSIVE_RESIGNATION", "MASKING", "TEMPORAL_GRIEVANCE", "EXCLUSION_CONTRAST", "IRONIC_DEFERENCE", "FAINT_PRAISE", "RETROSPECTIVE_HOPE") and state_v > CENTER:
excess = state_v - CENTER
pull = sm.v_weight * sm.confidence * FORCE_SCALE * (1.0 + excess / 50.0)
state_v += pull
if sm.pattern == "RHETORICAL_SELF_NEGATION" and state_w > CENTER:
w_excess = state_w - CENTER
w_pull = sm.w_weight * sm.confidence * FORCE_SCALE * (1.0 + w_excess / 50.0)
state_w += w_pull
sm = StructureMatch(
pattern=sm.pattern, confidence=sm.confidence,
matched_indices=sm.matched_indices, description=sm.description,
v_weight=sm.v_weight, d_weight=sm.d_weight, u_weight=sm.u_weight,
g_weight=sm.g_weight, w_weight=0.0,
)
elif sm.pattern == "CHOPPER_SPLIT" and sm.matched_indices:
chop_pos = sm.matched_indices[0]
after_words = [wr for wr in roles if wr.position > chop_pos]
after_v_sum = 0
for wr in after_words:
wf = wr.force or VOCABULARY.get(wr.word)
if wf:
after_v_sum += wf[0]
has_negator_after = any(wr.role == "NEGATOR" for wr in after_words)
if (state_v > CENTER and (after_v_sum < 0 or (after_v_sum == 0 and has_negator_after))):
distance = state_v - CENTER
state_v -= distance * 1.5 * sm.confidence
elif (state_v < CENTER and after_v_sum > 10):
distance = CENTER - state_v
state_v += distance * 0.4 * sm.confidence
else:
state_v += sm.v_weight * sm.confidence * FORCE_SCALE
state_a += sm.a_weight * sm.confidence * FORCE_SCALE
state_d += sm.d_weight * sm.confidence * FORCE_SCALE
state_u += sm.u_weight * sm.confidence * FORCE_SCALE
state_g += sm.g_weight * sm.confidence * FORCE_SCALE
state_w += sm.w_weight * sm.confidence * FORCE_SCALE
# Compound bond event anchoring (Council Round 8)
# If a compound bond exists, it's the EVENT NUCLEUS — anchor V toward its charge
_COMPOUND_VOCAB_KEYS = set(_COMPOUND_BONDS.values()) | set(_COMPOUND_BONDS_TRI.values())
for wr in roles:
if wr.word in _COMPOUND_VOCAB_KEYS:
f = wr.force or VOCABULARY.get(wr.word)
if f and abs(f[0]) >= 25:
# Anchor: push V toward the compound's charge direction
anchor_push = f[0] * 0.4 * FORCE_SCALE
state_v += anchor_push
# Sarcasm inversion penalty (from interpret_context step 7)
sarcasm_penalty = context.get("sarcasm_penalty", 0.0)
if sarcasm_penalty != 0.0:
state_v += sarcasm_penalty * FORCE_SCALE
context["state_v"] = state_v
context["state_a"] = state_a
context["state_d"] = state_d
context["state_u"] = state_u
context["state_g"] = state_g
context["state_w"] = state_w
return context
# ── Stage 5b: Static Friction (Council Round 7) ──────────────────
# Gemini's model: if no atom has |dV| > threshold, the sentence lacks
# emotional conviction. The pendulum can't overcome static friction.
# V deviation from center is dampened by (|dV_max| / threshold)².
STATIC_FRICTION_THRESHOLD = 20 # minimum |dV| to overcome friction (15 was too aggressive)
def apply_static_friction(context: dict) -> dict:
"""Prevent weak negative/positive drift from accumulated noise.
If no word in the sentence has |dV| > threshold, the total V deviation
is dampened by a squared ratio. Stronger max atoms = less dampening.
Exempted when crisis structures fire (dangling bonds etc. have zero-charge
atoms but structural signals that override).
"""
state_v = context["state_v"]
roles = context.get("roles", [])
structures = context.get("structures", [])
# Find max absolute dV in the sentence
max_abs_dv = 0
for wr in roles:
f = wr.force or VOCABULARY.get(wr.word)
if f:
max_abs_dv = max(max_abs_dv, abs(f[0]))
# Exempt if crisis structures fired
_CRISIS_STRUCTS = {"DANGLING_BOND", "FAREWELL", "MASKING", "RESIGNATION",
"WORLD_CONTINUES", "FINALITY", "METHOD_ACQUISITION",
"SELF_REMOVAL", "SUSPICIOUS_CALM", "SELF_HARM_INTENT",
"EXHAUSTION", "NO_EXIT"}
has_crisis_struct = any(sm.pattern in _CRISIS_STRUCTS for sm in structures)
if max_abs_dv < STATIC_FRICTION_THRESHOLD and not has_crisis_struct:
# Squared ratio: smooth transition, harder to move with weaker atoms
friction = (max_abs_dv / STATIC_FRICTION_THRESHOLD) ** 2
deviation = state_v - CENTER
state_v = CENTER + deviation * friction
context["state_v"] = state_v
return context
# ── Stage 6: Apply W Coefficient ────────────────────────────────
def apply_w_coefficient(context: dict) -> dict:
"""Self-worth modulates valence via asymmetric exponential coupling.
Low W amplifies negative V (validation of broken state).
Low W suppresses positive V (rejection of contradictory energy).
At W=128 (neutral), negatives are amplified 1.5x (calibrated
negativity bias — the whole benchmark suite is fitted around it;
making neutral-W a no-op was tested 2026-06-11 and regressed
stress 271→263, crisis recall 49→46). At W=50, ~1.8x (capped).
GPT's equation:
w = (W - 128) / 128 → normalized: -1 (broken) to +1 (strong)
V_neg' = V_neg * (1 + β * e^(-w)) → amplify negatives when w < 0
V_pos' = V_pos * (1 - γ * (1 - e^w)) → suppress positives when w < 0
Reads: context["state_v"], context["state_w"]
Writes: context["state_v"]
"""
state_v = context["state_v"]
state_w = context["state_w"]
# Normalized self-worth: -1 = broken, 0 = neutral, +1 = strong
w_norm = (state_w - CENTER) / CENTER
W_BETA = 0.5 # negative amplification strength (was 0.8, too aggressive)
W_GAMMA = 0.3 # positive suppression strength (was 0.6, killed slang positives)
W_CAP = 1.8 # max amplification
displacement = state_v - CENTER
if displacement < 0:
# Negative V: amplify when W is low (w_norm < 0 → e^(-w_norm) > 1)
amp = min(W_CAP, 1.0 + W_BETA * exp(-w_norm))
state_v = CENTER + displacement * amp
elif displacement > 0:
# Positive V: suppress when W is low (w_norm < 0 → e^(w_norm) < 1)
sup = max(0.0, 1.0 - W_GAMMA * (1.0 - exp(w_norm)))
state_v = CENTER + displacement * sup
context["state_v"] = state_v
return context
# ── Stage 7: Apply Personality ──────────────────────────────────
def apply_personality(context: dict) -> dict:
"""Scale state by personality vector if provided.
Reads: context["state_*"], context.get("personality")
Writes: context["state_*"]
"""
personality = context.get("personality")
if personality is None:
return context
state_v = context["state_v"]
state_a = context["state_a"]
state_d = context["state_d"]
state_u = context["state_u"]
state_g = context["state_g"]
state_w = context["state_w"]
sensitivity = personality.emotional_sensitivity
state_v = CENTER + (state_v - CENTER) * sensitivity
state_a = CENTER + (state_a - CENTER) * sensitivity
state_d = CENTER + (state_d - CENTER) * sensitivity + personality.dominance_baseline
state_u = state_u * sensitivity
state_g = CENTER + (state_g - CENTER) * sensitivity + personality.gravity_bias
state_w = CENTER + (state_w - CENTER) * sensitivity
context["state_v"] = state_v
context["state_a"] = state_a
context["state_d"] = state_d
context["state_u"] = state_u
context["state_g"] = state_g
context["state_w"] = state_w
return context
# ── Stage 8: Saturate and Clamp ─────────────────────────────────
def saturate_and_clamp(context: dict) -> dict:
"""Tanh saturation, intent computation, and 0-255 clamping.
Reads: context["state_*"], context["force_flow"], context["roles"],
context["trace_entries"], context["structures"], context["words"]
Writes: context["vadug"], context["meta"]
"""
state_v = context["state_v"]
state_a = context["state_a"]
state_d = context["state_d"]
state_u = context["state_u"]
state_g = context["state_g"]
state_w = context["state_w"]
force_flow = context.get("force_flow")
roles = context.get("roles", [])
# Intent computation
state_i = compute_intent(force_flow, roles)
# Tier-1 passive aggression lifts I toward CONTROL: the speaker is
# fighting, deniably — the embedded grievance is a directed jab, not
# neutral information. Only lift out of the neutral/connect band; a
# strong directional reading (withdraw/attack) is never overridden.
_PA_CONTROL_PATTERNS = {
"TEMPORAL_GRIEVANCE", "EXCLUSION_CONTRAST",
"IRONIC_DEFERENCE", "FAINT_PRAISE", "RETROSPECTIVE_HOPE",
}
structures = context.get("structures", [])
if any(sm.pattern in _PA_CONTROL_PATTERNS for sm in structures) and \
90 <= state_i < 168:
state_i = 168
# Tanh saturation
state_v = CENTER + SATURATION * tanh((state_v - CENTER) / SATURATION)
state_a = CENTER + SATURATION * tanh((state_a - CENTER) / SATURATION)
state_d = CENTER + SATURATION * tanh((state_d - CENTER) / SATURATION)
state_u = SATURATION * tanh(state_u / SATURATION)
state_g = CENTER + SATURATION * tanh((state_g - CENTER) / SATURATION)
state_w = CENTER + SATURATION * tanh((state_w - CENTER) / SATURATION)
# Clamp to 0-255
result = VADUG(
v=int(round(max(0, min(255, state_v)))),
a=int(round(max(0, min(255, state_a)))),
d=int(round(max(0, min(255, state_d)))),
u=int(round(max(0, min(255, state_u)))),
g=int(round(max(0, min(255, state_g)))),
w=int(round(max(0, min(255, state_w)))),
i=state_i,
)
trace_dict = {
"trace": context.get("trace_entries", []),
"structures": context.get("structures", []),
"force_flow": force_flow,
"word_count": len(context.get("words", [])),
}
context["vadug"] = result
context["meta"] = trace_dict
return context
# ── Pipeline ────────────────────────────────────────────────────
class Pipeline:
"""Chainable pipeline of stage functions.
Each stage is a function(context: dict) -> dict.
Stages read from and write to the shared context dict.
"""
def __init__(self, stages=None):
self.stages = stages if stages is not None else self.default_stages()
@staticmethod
def default_stages():
return [
tokenize,
classify,
interpret_context, # V8.1: role reinterpretation before forces
compute_coefficients,
accumulate_forces,
apply_structures,
# apply_static_friction, # Council R7: needs higher threshold tuning, disabled for now
apply_w_coefficient,
apply_personality,
saturate_and_clamp,
]
def run(self, text: str, perspective: str = "speaker",
personality: Optional[PersonalityVector] = None) -> Tuple[VADUG, dict]:
"""Run the full pipeline on text.
Returns (VADUG, trace_dict) — same interface as compute_vadug().
"""
context = {
"text": text,
"perspective": perspective,
"personality": personality,
}
# Early exit for empty text
words = text.split()
if not words:
return VADUG(), {"trace": [], "structures": [], "word_count": 0}
for stage in self.stages:
context = stage(context)
return context["vadug"], context["meta"]
# ── Main entry point (thin wrapper) ────────────────────────────
def compute_vadug(
text: str,
personality: Optional[PersonalityVector] = None,
perspective: str = "speaker",
) -> Tuple[VADUG, dict]:
"""Compute VADUGWI coordinates for a text string.
Pipeline:
1. Split text into words
2. Layer 1: classify_sentence() -- structural roles
3. Layer 2: proximity_coefficient() -- distance-based influence
4. Layer 3: StructureDetector().detect_all() -- chess-like patterns
5. Physics loop: momentum + force blending
6. Structure adjustments
7. Personality adjustments (if provided)
8. Clamp to 0-255
perspective controls whose emotional state is being scored:
- "speaker": default. "I" = self, "you" = other. Scores the speaker.
- "listener": "you" = self, "I" = other. Scores the person being spoken to.
- "bystander": no self. "I" and "you" are both other people. Scores a
detached observer who sees the emotional content but takes no directed hits.
Returns (VADUG, trace_dict) where trace_dict contains:
- trace: list of per-word entries {word, role, coeff, v, a, d, u, g}
- structures: list of detected StructureMatch objects
- word_count: int
"""
return Pipeline().run(text, perspective=perspective, personality=personality)